Can you explain the concept "tuned" in detail ?

[moonlight1776]

Hello, xwzhong
I am confused about the concept "tuned" , how to use the matrix and how we can get it ?
tuned：For adaptation, the second configuration uses an additional transformation matrix of the sentence embeddings. The matrix is parameterized using the STS training data。
Thank you very much !

[xwzhong]

hi,
here is my opionion:
tuned means add a maxtric calculation after encode the sentence embedding( something like feedforward operation), that convert vector u(or v) to u*(or v*), then use formula 2 get similarity between u* and v*. using STS training data, we can optimize the matrix.
for the reason why the author do that, i guess, he want fit the STS similary degree ranging from 0 to 5. if we just use formula 2, it just a liner transformation, but the truth may be not, so, maybe that is why the author add the transformation matrix, for fitting the STS data itself.